Global Explanations for Multivariate Time Series Forecasting Models via $K$-Order Markov Approximations

Time-series forecasting needs specialized explanations: Most existing explainable AI methods weren't designed for forecasting models and incorrectly assume that data points are independent of each other, ignoring the sequential nature of time-series data. KARMA addresses this gap by building explanations that respect how events depend on their history.

The method identifies how much past data matters: KARMA determines the minimal history length (K) that a model actually needs to make predictions, which helps practitioners understand whether a forecasting model is appropriately using recent versus distant information—a critical question for responsible deployment.

Explanations can be organized hierarchically: The paper demonstrates a five-level explanation hierarchy derived from the model's learned patterns, allowing stakeholders to understand predictions at different levels of detail, from high-level trends (like in Beijing PM 2.5 forecasting) down to specific temporal dependencies.

KARMA recovers causal relationships better than alternatives: In controlled experiments with synthetic data containing known causal structures, KARMA successfully identified the true causal edges that models learned, while established methods like TimeSHAP performed worse—suggesting this approach provides more trustworthy explanations for time-dependent systems.

Real-world applicability demonstrated: The method was tested on actual weather pollution data (Beijing PM 2.5), showing it can generate practical, interpretable explanations for real forecasting systems rather than just theoretical examples.